The quarter's marquee earnings landed after Wednesday's close, and for the first time in this build-out cycle the market drew a sharp line between the two companies spending the most. Microsoft was rewarded and Meta was punished — not because one is investing and the other isn't, but because only one of them could point to the revenue the investment is supposed to produce. For European buyers weighing where their inference capacity sits, the split is a useful reminder that the money answering the AI question is still being spent almost entirely on US soil.
Microsoft's fiscal Q4 (the quarter ended 30 June) beat on nearly every line: revenue of roughly $90 billion, up about 18%, and net income of $35.77 billion versus $27.23 billion a year earlier. The number that moved the stock was Azure, which grew 43% — an acceleration from the prior quarter's 40% — with management guiding to about 45% constant-currency growth for the current quarter and noting that Azure had crossed $100 billion in annual revenue for the first time. Against roughly $190 billion of committed fiscal-2026 capital spending, that is the payoff investors had been waiting to see, and shares rose after hours.
Meta told the opposite story on the same afternoon. Revenue rose 28% to $60.8 billion, but earnings missed, quarterly capital expenditure hit $31.1 billion (more than double a year earlier), and free cash flow collapsed to under $1 billion as the spend outran the return. The stock fell nearly 10% after hours. The lesson for anyone building an AI strategy is not that the capex is reckless — both companies are still guiding it higher — but that the market has started grading it on delivered revenue rather than ambition. The capacity being funded, and rewarded, remains overwhelmingly hyperscale and US-jurisdiction; the question for regulated European workloads is less whether the build-out happens than under whose law the resulting servers run.
The Plumbing Under the Agents Just Got a Major Revision
While the money story dominated headlines, the more durable news for anyone actually shipping AI was quieter: the Model Context Protocol published its 2026-07-28 revision this week, the biggest change to the agent-tooling standard since it appeared. The headline shift is a move from a stateful, bidirectional protocol to a stateless request/response core, which lets an MCP server run behind an ordinary load balancer on serverless or edge infrastructure instead of holding a long-lived connection per client. Alongside it come a formal extensions framework, first-class support for long-running Tasks and interactive Apps, and — most relevant to enterprises — authorization hardening that aligns the protocol with standard OAuth practice (issuer validation and client-metadata documents).
That last piece is why this matters beyond developer ergonomics. As agent tooling becomes something a compliance team has to reason about, “how does this server authenticate, and can it run inside our own network” stops being a footnote. Anthropic has already said it is bringing the new spec to Claude, and deprecated features get at least a twelve-month runway — a rare instance of an AI standard treating backwards compatibility as a feature rather than an afterthought. For platforms that expose tools over MCP, the stateless core and the OAuth alignment are the difference between a demo integration and one an enterprise security review will actually pass.
Who Governs the Agents After You Deploy Them
The third front is control, and this week it drew real money. Tel Aviv-based Hush Security raised a $30 million Series A — with Akamai joining as a strategic investor alongside Battery Ventures and YL Ventures — to govern the identities of AI agents, letting enterprises register agents, strip out standing credentials, grant scoped just-in-time access, log activity, and revoke it from one place. The raise lands on a genuinely uncomfortable statistic: non-human identities already outnumber humans by roughly 45 to 1 in the average enterprise, and surveys this month found the overwhelming majority of organisations running agents while saying their existing identity tooling cannot manage them.
The gap between deployment and governance is the recurring theme of enterprise AI in 2026, and it is exactly where a sovereignty argument becomes an operational one rather than a philosophical one. An agent is a privileged actor with credentials, a purpose, and an access scope; when it acts on regulated data, the questions of who can see its activity, where the logs live, and under whose jurisdiction the access is adjudicated are not abstractions. Gartner's projection that the average large enterprise will run six figures of AI agents within two years is the reason this was a stealth-stage curiosity a year ago and a funded category now.
Quick Hits
- EU AI Act enforcement powers go live 2 August — Three days out, the Commission's GPAI enforcement machinery activates: Article 101 penalties of up to €15 million or 3% of global turnover, plus powers to demand documentation, run evaluations, and restrict models from the EU market. The obligations have applied since August 2025; it is the enforcement that is new.
- 1,100+ lab employees ask Washington for a “pacing” mechanism — An open letter signed by staff across OpenAI, Anthropic, Google, and Meta urges the US government to build verifiable tools that could coordinate a slowdown if capability outruns oversight. It stops short of calling for a pause.
- Mistral's open-weight MoE stays in early access — The “fat but sparse” mixture-of-experts model Arthur Mensch has been teasing is still partner-only, with no parameter count, benchmarks, or licence disclosed. Europe's clearest open-frontier candidate remains a promise rather than a download.
- Gemini 3.5 Pro slips again — Google's flagship is still in partner testing over coding shortfalls, with no new timeline, even as pre-training on Gemini 4 is reported to have begun.
- Amazon reports Friday — AWS growth is the next hyperscaler data point after the Microsoft–Meta split, and the clearest read yet on whether cloud demand is keeping pace with the industry's capital spending.